Skip to contents

Agent Skills

Source: vignettes/agent-skills.qmd

Overview

Agent Skills are pre-packaged capabilities that extend Claude’s abilities for specialized tasks. This vignette covers:

  1. Pre-built skills: PowerPoint, Excel, Word, PDF
  2. Local skills: Load from filesystem
  3. Uploaded skills: Persist on Anthropic servers
  4. Custom skills: Create your own

Pre-Built Skills

Anthropic provides 4 pre-built skills for document processing:

library(artclaude)

# List pre-built skills
claude_skill_list(type = "prebuilt")
#   skill_id name       type
# 1 pptx     PowerPoint prebuilt
# 2 xlsx     Excel      prebuilt
# 3 docx     Word       prebuilt
# 4 pdf      PDF        prebuilt

PowerPoint Skill

Create and modify PowerPoint presentations:

chat <- claude_new(
  skills = list(claude_skill("pptx")),
  tools = list(claude_code_exec()),
  beta = c(beta_headers$BETA_SKILLS, beta_headers$BETA_CODE_EXEC_BASH)
)

chat$chat("Create a presentation about R package development with 5 slides:
  1. Title slide
  2. Why R packages
  3. Package structure
  4. Documentation with roxygen2
  5. Testing with testthat")

Excel Skill

Create and modify Excel spreadsheets:

chat <- claude_new(
  skills = list(claude_skill("xlsx")),
  tools = list(claude_code_exec()),
  beta = c(beta_headers$BETA_SKILLS, beta_headers$BETA_CODE_EXEC_BASH)
)

chat$chat("Create an Excel file with:
  - Sheet 1: Monthly sales data for 2024
  - Sheet 2: Pivot table summarizing by quarter
  - Sheet 3: Chart showing trend")

Word Skill

Create and modify Word documents:

chat <- claude_new(
  skills = list(claude_skill("docx")),
  tools = list(claude_code_exec()),
  beta = c(beta_headers$BETA_SKILLS, beta_headers$BETA_CODE_EXEC_BASH)
)

chat$chat("Create a Word document: Technical Specification for API v2.0
  Include: Overview, Endpoints, Authentication, Rate Limits, Examples")

PDF Skill

Extract and process PDF content:

chat <- claude_new(
  skills = list(claude_skill("pdf")),
  tools = list(claude_code_exec()),
  beta = c(beta_headers$BETA_SKILLS, beta_headers$BETA_CODE_EXEC_BASH)
)

# Upload PDF via Files API
file <- claude_file_upload("annual_report.pdf")

chat$chat(
  claude_file_content(file$id),
  "Extract all financial tables from this PDF"
)

Local Skills

Creating a Skill

Skills follow the agentskills.io specification:

my_skill/
+-- SKILL.md      # Required: YAML frontmatter + instructions
+-- templates/    # Optional: template files
+-- scripts/      # Optional: helper scripts

SKILL.md format:

---
name: greeting
description: Greet users professionally
---

Always greet users with "Welcome to Artalytics!" followed by
a brief introduction to the platform's capabilities.

Using Local Skills 

library(artclaude)

# Create skill directory
skill_dir <- tempfile("skill_")
fs::dir_create(skill_dir)

writeLines(c(
  "---",
  "name: greeting",
  "description: Greet users professionally",
  "---",
  "",
  "Always greet users with 'Welcome to Artalytics!'",
  "followed by a brief introduction."
), fs::path(skill_dir, "SKILL.md"))

# Load and use
chat <- claude_new(
  skills = list(claude_skill_local(skill_dir)),
  tools = list(claude_code_exec()),
  beta = c(beta_headers$BETA_SKILLS, beta_headers$BETA_CODE_EXEC_BASH)
)

chat$chat("Please greet me")

Skills with Resources

Add template files and scripts:

skill_dir <- "~/skills/report_generator"
fs::dir_create(fs::path(skill_dir, "templates"))

# Main skill file
writeLines(c(
  "---",
  "name: report_generator",
  "description: Generate formatted reports",
  "---",
  "",
  "Use the template in templates/report.md to create reports.",
  "Follow the structure and formatting guidelines."
), fs::path(skill_dir, "SKILL.md"))

# Template file
writeLines(c(
  "# {title}",
  "",
  "## Executive Summary",
  "{summary}",
  "",
  "## Findings",
  "{findings}"
), fs::path(skill_dir, "templates", "report.md"))

# Use skill
chat <- claude_new(
  skills = list(claude_skill_local(skill_dir)),
  tools = list(claude_code_exec()),
  beta = c(beta_headers$BETA_SKILLS, beta_headers$BETA_CODE_EXEC_BASH)
)

chat$chat("Create a report about Q4 sales using the template")

Uploaded Skills

Upload for Persistent Use

Upload skills to Anthropic’s servers:

# Create skill
skill_dir <- tempfile("skill_")
fs::dir_create(skill_dir)
writeLines(c(
  "---",
  "name: code_reviewer",
  "description: Review R code for best practices",
  "---",
  "",
  "Review R code for:",
  "- Proper use of data.table",
  "- Native pipe usage",
  "- Package qualification (pkg::function)",
  "- Documentation completeness"
), fs::path(skill_dir, "SKILL.md"))

# Upload
result <- claude_skill_upload(skill_dir)
result$skill_id # "skill_abc123..."

# Use in future sessions
chat <- claude_new(
  skills = list(claude_skill(result$skill_id)),
  tools = list(claude_code_exec()),
  beta = c(beta_headers$BETA_SKILLS, beta_headers$BETA_CODE_EXEC_BASH)
)

chat$chat("Review this function: get_data <- function() { ... }")

Listing Skills 

# List all skills
all_skills <- claude_skill_list(type = "all")

# List only pre-built
prebuilt <- claude_skill_list(type = "prebuilt")

# List only uploaded
uploaded <- claude_skill_list(type = "uploaded")

Skill Design Best Practices

Clear Instructions 

# Good: Specific, actionable
"When analyzing code:
1. Check for data.table usage
2. Verify pipe is native |>
3. Ensure external calls use pkg::function()
4. Confirm roxygen docs are complete"

# Not ideal: Vague
"Review code for quality"

Contextual Information

Provide context that helps Claude succeed:

---
name: art_evaluator
description: Evaluate artwork descriptions
---

You are an art expert specializing in Impressionist and Post-Impressionist
movements. When evaluating artwork descriptions:

1. Verify historical accuracy
2. Check artist attribution
3. Assess style classification
4. Validate date ranges
5. Review medium/technique descriptions

Reference established art history sources and note any discrepancies.

Resource Organization 

artwork_skill/
+-- SKILL.md
+-- templates/
|   +-- evaluation_template.md
|   +-- report_template.md
+-- reference/
|   +-- impressionist_artists.csv
|   +-- movement_timeline.json

Use Cases

Domain-Specific Agent 

# Art curation assistant
art_skill <- claude_skill_local("~/skills/art_curator")

chat <- claude_new(
  skills = list(art_skill),
  tools = list(claude_code_exec(), claude_web_search()),
  beta = c(beta_headers$BETA_SKILLS, beta_headers$BETA_CODE_EXEC_BASH)
)

chat$chat("Curate a collection of Impressionist works from 1870-1890")

Workflow Automation 

# Document processing pipeline
skills <- list(
  claude_skill("pdf"),
  claude_skill("xlsx"),
  claude_skill("pptx")
)

chat <- claude_new(
  skills = skills,
  tools = list(claude_code_exec()),
  beta = c(beta_headers$BETA_SKILLS, beta_headers$BETA_CODE_EXEC_BASH)
)

chat$chat("Extract data from report.pdf, create Excel summary, generate PowerPoint")

Limits

  • Maximum 8 skills per request
  • Skills require code execution tool
  • Skills require appropriate beta headers

API Reference

Skills Guide: https://platform.claude.com/docs/en/build-with-claude/skills-guide